Ask Your Question

face recognition and skin or similarity problems

asked 2016-07-13 07:39:27 -0500

atv gravatar image

This is a followup of my previous question but:

I have 4 trained persons in my database (200 pics each). 1 is of colored (brown) skin type.

If i show a unknown person of similar skin type, it matches, with the same accuracy as a regular good match (no elevation in eigenvalue, so threshold won't do much good. I get the same when i have an trained old person in my database, it will match against a picture shown of a ..old person.

How do i protect against this? Do i need to expand my own trained database more get an increase in eigenvalues? Or do i need to resort to more advanced recognition techniques?

Thank you for reading!

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2016-07-13 07:43:04 -0500

MRDaniel gravatar image

updated 2016-07-13 07:44:12 -0500

You need more negative samples of faces.

Create a new class of "Unkown". This will then make the matcher return a possible Unknown as the features become more discriminant.

Essentially, your model is too simple.

It's like saying, heads or tails. (Well, it landed on the edge of the coin...... may as well say it's tails).

edit flag offensive delete link more


So expand my own database. So i create new label of a dummy person and add lots of pictures of people that have no bearing on the people i want to recognize. Is that correct?

Should i add lots of different new persons with their own picture set or just 1 with a giant set of negative images.

atv gravatar imageatv ( 2016-07-13 07:52:10 -0500 )edit

I would try both. I think that you should have just one dummy category. The problem there is, you may add someone who later needs to be put into a positive category. This is the dude for Face Recognition if you have not come across this article yet.....

MRDaniel gravatar imageMRDaniel ( 2016-07-13 07:55:59 -0500 )edit
Login/Signup to Answer

Question Tools

1 follower


Asked: 2016-07-13 07:39:27 -0500

Seen: 206 times

Last updated: Jul 13 '16